Parallel algorithms for Scientific Computing
This course is organized by the Division of Scientific Computing at the IT Department (DCS)
If you intend to participate in the course, please send an email to maya.neytcheva@it.uu.se. Thank you.
General description
The course focuses on parallel algorithms for scientific computing problems (not on parallel programming). We will cover parallel algorithms for a range of problem classes in Scientific Computing, from linear algebra (with a focus on sparse problems related to solving PDE), iterative solvers, multigrid, particle interaction problems (Barnes-Hut, multipole etc) to algorithms, typical for data intensive applications and machine learning.
The material is split into three parts as follows.
- Parallelization issues in numerical linear algebra (focus on PDE-related problems)
- Fast multipole methods and more
- Algorithms for data intensive applications and machine learning
Examination and grading
The course gives 5 hp.
To pass the course, one has to fulfill an assignment for each part, to be delivered in a written form.
It is possible to obtain 7.5 hp by doing an additional project that includes some testing on a parallel computer facility (to be discussed individually).
Course prerequisites
Experience of scientific computing problems and basic knowledge of scientific computing algorithms. Detailed knowledge of algorithms like conjugate gradients, multigrid and multipole approximations is not needed. Experience of programming for scientific computing in languages like C, but you do not need to have taken a course in parallel programming to do the assignments. An optional project work (for obtaining 7,5 hp) should contain implementation and experiments using a parallel system, thus, presumably more experience of parallel programming will be needed.
The lectures will be given at the Department of Information Technology, Uppsala University.
Lecturers
Sverker Holmgren (SH)
Carl Nettelblad (CN)
Maya Neytcheva (MN)
Additional course materials (to be added during the course)
Part 1: Assignment
Part 2:
Part 3:
Detailed time schedule
| Date |
Topic(s) |
Time |
Location |
Lecturer |
| Nov 28 |
Lecture 1: Sparse matrices and storage schemes. |
13:15-15:00 |
2414b |
MN |
| Nov 30 |
Lecture 2: Iterative methods and parallelization. Pipelined versions of CG and GMRES - an attempt to achieve better scaling for Very Large/Exascale |
10:15-12:00 |
2415b |
MN |
| Dec 6 |
Lecture 3: Preconditioners - parallelization aspects
|
13:15-15:00 |
1406 |
MN |
| Dec 7 |
Parallelization of incomplete factorization methods. Talk by prof. Thomas Huckle at the TDB seminar
|
13:15-14:00 |
2344 |
MN |
| | | | | |
| XX |
Lecture 1: TBA |
XX |
XXXX |
SH |
| XX |
Lecture 2: TBA
|
XX |
XXXX |
SH |
| XX |
Lecture 3: TBA
|
XX |
XXXX |
SH |
| | | | | |
| XX |
Lecture 1: TBA |
XX |
XXXX |
CN |
| XX |
Lecture 2: TBA
|
XX |
XXXX |
CN |
| XX |
Lecture 3: TBA
|
XX |
XXXX |
CN |
Recommended books:
- Part 1:
- Data storage schemes for parallel computations
old source, 1992,
new source, 2017
- P. Ghysels, W. Vanroose, Hiding global synchronization latency in the preconditioned conjugate
gradient algorithm, Parallel Computing, 40 (2014), pp. 224-238,
(also here)
- P. Sanan, S.M. Schnepp, D.A. May, Pipelined flexible Krylov subspace methods
- Pipelined solvers, performance comparison
- PETSc: pipelined CG , PETSc: pipelined GMRES
- Part 2: To come
- Part 3: To come
Last changed on Nov 22, 2017.
Mail to: Maya dot
Neytcheva "at"
it dot
uu dot se "